The Benefit of Collective Intelligence in Community-Based Content Moderation is Limited by Overt Political Signalling

Technical Summary: The Benefit of Collective Intelligence in Community-Based Content Moderation is Limited by Overt Political Signalling

Problem Statement
Social media platforms face significant challenges regarding the rapid dissemination of misinformation and misleading content. While traditional expert fact-checking lacks scalability and automated algorithms suffer from training data biases, platforms have increasingly adopted community-based moderation systems (e.g., X's Community Notes, Meta's Community Notes, TikTok's Footnotes). These systems rely on users to write contextual notes and rate them for helpfulness, with the goal of displaying notes that receive positive ratings from an ideologically diverse set of raters.

However, current implementations face critical limitations. Research indicates that the existing design, which relies on individually authored notes rated by others, is highly sensitive to rater bias and in-group preferences. Simulation and observational studies suggest that partisanship heavily influences evaluations, with users disproportionately flagging counter-partisan content and rating co-partisan notes as more helpful. Furthermore, the process is often slow, with a significant lag between post publication and note display. A key theoretical gap exists regarding whether moving from an aggregation of independent contributions to explicit Collective Intelligence (CI)—where users collaborate directly to write notes—can improve quality, and how political diversity and identity signalling within these collaborative groups affect outcomes.

Methodology
The authors conducted a controlled online experiment involving 432 participants recruited from Prolific, self-identifying as Democrats or Republicans. The study utilized a 2x3 factorial design to test the effects of collaboration, team composition, and identity signalling:

1. Task Flow: Participants first wrote individual notes providing context for 40 politically misleading posts (sourced from Democrat and Republican accounts). Subsequently, they were paired into teams to communicate via a chat interface and revise their notes into a single, joint version.
2. Team Configurations: Participants were randomly paired to form three group types:
   • Democrat–Democrat (DD)
   • Republican–Republican (RR)
   • Democrat–Republican (DR)
3. Treatment Conditions: Teams were randomly assigned to one of two conditions:
   • Overt: Participants were explicitly informed of their partner's political affiliation (e.g., nicknames like "Democrat 1").
   • Covert: Political affiliations were not disclosed (neutral nicknames).
4. Evaluation: To assess note quality, the researchers recruited 1,610 additional Prolific participants (805 Democrats, 805 Republicans) to rate the helpfulness of both individual and collaborative notes on a 0–10 scale. Additionally, a panel of three non-US expert PhD students provided neutral evaluations.
5. Metrics: The study measured the change in helpfulness ($I$) between collaborative notes ($H_{ab}$) and the average of the original individual notes ($H_a, H_b$) across Democrat raters ($I_D$), Republican raters ($I_R$), and experts ($I_E$). Statistical analysis included parametric t-tests, non-parametric bootstrap resampling, and Ordinary Least Squares (OLS) regression controlling for note length and link count.

Key Results
The study yielded three primary findings corresponding to the authors' hypotheses:

• H1: Collaboration Improves Quality. Collaboratively written notes were rated as significantly more helpful than individually written notes by experts (mean increase of ~17%; $p < 0.01$). While the effect size was smaller in crowdsourced ratings, it remained statistically significant for Republican raters.
• H2: Asymmetric Impact of Diversity. The benefit of political diversity (DR teams) was not uniform. For Republican posts, politically diverse teams produced significantly better notes than homogeneous Republican teams. However, for Democrat posts, team composition (DD, RR, or DR) did not meaningfully affect the improvement in note quality. This suggests an asymmetry in how political diversity functions depending on the target content's source.
• H3: The Cost of Overt Signalling. The benefits of collaboration were significantly diminished when political identities were overtly disclosed. In the covert condition, collaborative notes showed a substantial increase in helpfulness. In the overt condition, this advantage nearly vanished, and in some cases, collaborative notes were rated as less helpful than individual ones. This negative effect persisted across expert, Democrat, and Republican ratings. Chat analysis indicated that overt signalling increased partisan identity salience, disagreement, and concessionary behavior, suggesting that visible identity cues introduce friction that undermines collective problem-solving.

Significance and Claims
The paper posits that while community-based content moderation holds promise, its effectiveness is contingent on the specific design of collaborative interactions. The authors claim that:

1. Collaboration is a viable mechanism for improving the quality of contextual notes, but it is not a standalone solution; the interactional environment is critical.
2. Political identity signalling acts as a constraint on collective intelligence. Making political affiliations explicit during the note-writing process triggers partisan friction, reducing the epistemic benefits of collaboration.
3. Design Implications: Platforms should consider designs that encourage exposure to diverse perspectives (e.g., algorithmic matching of users with different inferred viewpoints) while preserving anonymity regarding political identity. This approach aims to retain the benefits of diverse input without triggering the stereotyping and distrust associated with overt identity cues.

The authors maintain a modest stance, acknowledging that their experimental setting (structured, incentivized, synchronous) differs from real-world asynchronous volunteer moderation. They also note that the study does not fully isolate collaboration from order effects, as participants always wrote individually before collaborating. Nevertheless, the findings highlight that the structure of collaboration and the management of social cues are as important as the diversity of the participants themselves in determining the success of community-based moderation systems.